Direct Motion Estimation via Instantaneous Doppler Measurements

40 concepts32 papers1 proposal

To understand how a robot or autonomous vehicle moves through space, traditional perception systems rely on temporal differencing. Whether using cameras (optical flow) or standard time-of-flight LiDAR (scan matching), the core recipe remains the same: capture a snapshot at time $t$, capture another at time $t+1$, find corresponding points between the two, and calculate velocity based on how far those points moved.

While intuitive, this approach has fundamental flaws. It introduces latency, demands heavy computational overhead for data association, and completely fails in "geometrically degenerate" environments—like long, featureless tunnels or highways—where frame $t$ and frame $t+1$ look identical.

A rapidly growing body of research bypasses temporal differencing entirely by leveraging instantaneous Doppler measurements. By utilizing Frequency-Modulated Continuous-Wave (FMCW) radar and LiDAR sensors, tracking systems can measure the exact radial velocity of every single return point in a single snapshot. This architectural shift fundamentally changes how machines estimate their own ego-motion and perceive dynamic objects in the environment.


The Geometry of Single-Scan Velocity

How do you resolve a full 3D or 2D velocity vector from a sensor that only measures 1D radial velocity (speed directly toward or away from the sensor)? The answer lies in the geometric distribution of environmental reflections.

When an FMCW sensor moves through a static environment, every stationary object it hits returns a signal shifted in frequency by the Doppler effect. For any given point $i$, the measured radial velocity $v_r$ is the projection of the sensor's true velocity vector $\mathbf{v}$ onto the unit direction vector $\mathbf{u}_i$ pointing toward that reflection:

$$v_r = \mathbf{v} \cdot \mathbf{u}_i$$

This is a single linear equation with three unknowns (the $x, y, z$ components of $\mathbf{v}$). If the sensor detects at least three non-collinear stationary points in a single scan, it can instantly solve for its full 3D velocity vector using a linear system.

   [Stationary Wall Point A] (Returns radial velocity v_rA)
          ^
          |   u_A
          |
       (Sensor) ------> True Ego-Velocity Vector (v)
          |
          |   u_B
          v
   [Stationary Wall Point B] (Returns radial velocity v_rB)

   By observing v_rA and v_rB along distinct angles u_A and u_B, 
   the true 2D/3D velocity vector (v) is resolved instantly.

In practice, environments contain hundreds of points, some of which belong to moving objects rather than the static background. To handle this, Andres Galeote-Luque et al. demonstrate that a vehicle's full 3D motion can be estimated from a single scan by applying RANSAC to filter out dynamic outliers and using weighted least-squares regression on the remaining static returns. Similarly, Mian Guo applies least-squares optimization directly to stationary environmental elements isolated via Doppler analysis, computing the vehicle's 3D motion vector without relying on any external odometry or frame-to-frame tracking.

Even when explicit point clouds are unavailable, the underlying frequency shifts can be exploited directly from raw data. Fraser Rennie et al. show that convolutional neural networks can be trained to regress ego-motion directly from the frequency-shift artifacts and alternating modulation patterns of a single FMCW radar scan, bypassing temporal correspondence entirely.


Bypassing Scan-Matching and Mitigating Drift

Because Doppler-based velocity estimation relies on frequency shifts rather than geometric landmarks, it is immune to the failures that plague traditional scan-matching algorithms (like the Iterative Closest Point algorithm).

In a featureless underground mine or a uniform concrete tunnel, a traditional LiDAR or radar system cannot determine if it has moved forward because the geometric shape of the environment remains unchanged. This causes severe longitudinal drift. However, the Doppler shift does not care if a wall is flat and featureless; as long as the surface scatters radio or light waves back to the sensor, the frequency shift provides an absolute velocity reference.

Vladimír Kubelka et al. show that simply integrating instantaneous Doppler radar velocities with an IMU outperforms complex 3D scan-matching pipelines in feature-sparse environments, achieving remarkably low localization drift without any data association. In the optical domain, Weitong Wu integrates FMCW LiDAR Doppler measurements with a continuous-time trajectory model, demonstrating that these instantaneous measurements act as an anchor that completely mitigates longitudinal drift in geometrically degenerate tunnels.

As noted in a comprehensive survey by Ayoung Kim, this hardware-level velocity measurement also allows systems to perform motion compensation and eliminate motion distortion—the warping of point clouds caused by sensor movement during a sweep—without needing external sensors or iterative alignment algorithms.


Instantaneous Segmentation and Tracking

The utility of Doppler data extends far beyond estimating the sensor's own ego-motion; it serves as a powerful mechanism for environmental perception and object tracking.

In traditional perception pipelines, separating moving objects (like pedestrians or cars) from the static background requires tracking clusters over multiple frames to see if their positions change. With Doppler-capable sensors, this separation happens instantly at the individual point level. If a point's measured radial velocity does not match the projected ego-velocity vector of the vehicle, that point is immediately flagged as dynamic.

Traditional Tracking:
[Frame t: Object at X] ----(Compute Displacement)----> [Frame t+1: Object at Y] -> Velocity calculated

Doppler Tracking:
[Single Snapshot: Object at X with frequency shift Δf] -> Velocity calculated instantly

Aldrich Zhijun Pan et al. exploit this by integrating radial relative velocity directly into a point feature encoder, allowing an end-to-end network to immediately differentiate dynamic objects from static background noise and drastically reduce false positives in complex environments. This instantaneous velocity signature is so discriminative that Changsong Pang et al. utilize it within a spatial-temporal transformer network to simultaneously segment moving objects and refine ego-velocity estimates under adverse weather conditions.

For downstream tracking and prediction, knowing an object's exact velocity vector from a single snapshot eliminates the latency of initializing a track. Bin Yang et al. show that combining 1D radial Doppler measurements with predicted object headings through an attention mechanism allows networks to derive highly accurate 2D velocity vectors. This approach bypasses multi-frame temporal analysis, a capability that Ramin Nabati et al. leverage to estimate precise object velocities from a single data snapshot.


Multi-Modal Sensor Fusion

While Doppler measurements provide unparalleled velocity data, radar sensors often suffer from lower spatial resolution compared to cameras or standard LiDAR. The frontier of this research direction focuses on fusing the instantaneous velocity cues of Doppler sensors with the rich spatial and semantic features of other modalities.

When cameras are paired with Doppler radar, the radar provides the explicit motion cues that vision lacks. Lukas Stäcker et al. and ZhiweiLin et al. demonstrate that injecting radar Doppler features into vision-based Bird's Eye View (BEV) detection networks slashes velocity estimation errors by more than half compared to relying on sequential camera frames.

This multi-modal synergy is particularly vital for safety-critical tasks like trajectory prediction and tracking. Meet Shah et al. show that fusing radar's instantaneous velocity with LiDAR and HD maps yields a 52% reduction in prediction error for accelerating or decelerating vehicles, because the model does not have to wait for sequential position changes to detect a change in acceleration. By mapping these direct Doppler measurements into unified representations, as explored by Liangyan Gui et al., modern autonomous systems achieve a level of all-weather motion awareness that position-only sensors simply cannot replicate.

Key Foundational Work

LiRaNet: End-to-End Trajectory Prediction using Spatio-Temporal Radar Fusion (2020) — contributed 1 concept elements to this direction.
Seeing Around Street Corners: Non-Line-of-Sight Detection and Tracking In-the-Wild Using Doppler Radar (2020) — contributed 1 concept elements to this direction.
Radar-only ego-motion estimation in difficult settings via graph matching (2019) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_radar_only_ego_motion_estimation_06118ce9,
  title = {radar-only ego-motion estimation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/radar-only-ego-motion-estimation/rdr_06118ce955d74b019f1597293a879cfd},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_radar_only_ego_motion_estimation_06118ce9_v1,
  title = {radar-only ego-motion estimation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/radar-only-ego-motion-estimation/rdr_06118ce955d74b019f1597293a879cfd/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake